Record a current-main affected-CI cost audit and close disproven diagnosis lanes

[gunbai-bot]

Summary

Terminal accounting PR for the CI-cost research arc. No new mechanism. Expected wall gain: zero.

Harvests one current-main production floor (run 30863019228, head ed9785817ed8) into gunbc.ci_cost_arc_closeout_receipt and updates five-minute / roadmap authorities so merged work is no longer labeled in-flight/NEXT/probable.

Settled honestly

Mechanism	Disposition
#7522 net vs 197.13s	Capture MEASURED 110820 ms. NET OPEN — stamp+gate stage walls absent from recoverable gh run --log. Trigger: upload on-success-stage-{1,2}-receipt.tsv.
#7671 fleet native	Enrollment MERGED. Fleet rate OPEN (known fallback:native_realization_refused on run 30767841790; no native_count in recoverable 30863019228 log). Not a completed native speedup.
#7534 after #7728	OOM fixed. Warm-hit skip still PartiallyDelivered (skip counter proof OPEN). Opt-in; does not bypass corpus/index startup.

Harvested walls (Subject A)

• build ≈ 237s · regen ≈ 510s · heal skipped · ci ≈ 2509s · pipeline sum ≈ 54.3 min
• selection SelectionApplied 297/916
• resolve 431s · eval 65s · peak_rss ~8.5 GiB · cgroup peak ~10.8 GiB
• assembly-split (additive): rewire_import_str ~165s dominant among named rows

Terminal table

warm admission MERGED (net OPEN) · native substrate MERGED · production native rate OPEN · exact-tree opt-in PartiallyDelivered · retention-vs-drain CLOSED REJECT · entry union CLOSED · selection honesty LANDED · dominant open cost = per-entry assembly

Subject B (full-corpus/falsifier) marked OPEN with trigger — not invented from truncated logs.

Test plan

• Receipt witness dag/test/claim/ci_cost_arc_closeout_receipt_witness_test.dag
• CI green on this PR
• Confirm no dual disposition: md + five_minute_ci_gate_program_note both drop retention PARKED / in-flight for merged steps

[gunbai-bot]

Addressed review 47900:

1. Duration surface — harvested_capture_wall / harvested_resolve_wall / harvested_eval_wall / cold_admission_baseline_wall now return Millisecond via millisecond(count:); witness compares with millisecond_count (unit-modeling wall per floor_component_receipt_unit_surface_note).
2. design_document open thread — rewritten so Root-cause disk-tier repeat-resolve memory growth (>8GiB OOM on second resolve through v1 disk cache seam) #7728 closes the OOM; PartiallyDelivered is pinned to the unenrolled skip-counter proof, not “reliably OOMs the CI runner.”

— sent from lively-crab-41

[gunbai-bot]

CI fix for run 30867533644 (heal + regen panic):

1. ci_cost_arc_closeout_receipt_witness_test.dag — witness_resolve_dominates_eval had a line-broken > comparison (expected expression, found Gt). Put the comparison on one line; claim-run greens locally.
2. DESIGN.md — synced disk-tier open thread to design_document.dag (addresses review 47901 drift).

— sent from lively-crab-41

[gunbai-bot]

Addressed operator request-changes on the closeout shape:

1. No prose String rows in gunbc.ci_cost_arc_closeout_receipt — only typed walls (Millisecond), counts (Nat), peaks (ByteSize), subject identity, and settle variants.
2. Remainder first-class — ci_job_wall / classified_ci_wall / unattributed_ci_wall with executed reconciliation (unattributed = 1902042); pipeline sum likewise.
3. Directional claim narrowed — end_to_end_residual=UnattributedCiJobWall; compiler_next_measured_target=PerEntryAssembly (not end-to-end dominant). Freeze entry-view until PR2.
4. HandHarvestedDatedAudit subject; full-corpus NotHarvested.
5. PR retitled to the audit framing.

— sent from lively-crab-41

[gunbai-bot]

Verified review 47922 against HEAD 3acf692f: verdict is APPROVE with no Findings. No code change required (dashboard “BLOCKING REVIEW” relay mismatch).

— sent from lively-crab-41